Before you start
- Source image: Upload one image to establish the video's starting frame. The interface has no uploaded default. The saved graph references two_characters.png; its presence in your environment is not established.
- Video prompt (optional): Text describing action, camera movement, scene changes and optional sounds, up to 2,000 characters. Although the prompt field is structurally required, an empty string is permitted.
Start with an image and return one video
This recipe returns one Generated video at port output:15. Start with your own image and the supplied Pro, 720p, six-second configuration; Generate audio is enabled initially. Provider authentication must already be configured, and generation uses provider billing.
The saved prompt describes a chase across three two-second shots, with dialogue, sound and color directions. Those instructions guide generation; the retained material does not establish an editing timeline or exact adherence to cuts and timestamps.
- Upload your starting image through Source image. This selects the image loaded by node 16; keep its existing connection to the generation node.
Look for: Node 16 supplies its IMAGE output at slot 0 to node 14 input image.
- Replace the chase prompt with action and camera instructions suited to your image, using at most 2,000 characters. Fit any timed shots within Duration. Include desired dialogue or sound effects if Generate audio is enabled, or leave the prompt empty if you want no text guidance.
Look for: The generation request combines your starting image with the supplied text. Exact adherence to the requested action, timing and speech remains unestablished.
Outputs: output:15
Prepare the provider integration
Configured provider authentication is required. Generation sends the image and generation request to Vidu and uses provider billing. Authentication setup, prices and account limits are not specified in the retained material.
The preparation record lists tea-time-labs/sweet-tea-nodes as a required node pack. It is a prerequisite resource, not an executable fallback recipe. The graph records comfy-core version 0.11.1 for LoadImage and Vidu3ImageToVideoNode; those metadata do not establish a complete installation compatibility range. No local model download is listed.
GPU capacity, system RAM, comparable run timing and supported platforms remain unspecified. The preparation record's 1024×1024 workload label does not establish a required image size or the video's exact pixel dimensions.
Generation reports missing provider authentication.
The Vidu generation integration requires configured provider authentication.
Check: Prompt, duration and seed inputs do not configure credentials.
Configure provider authentication in the execution environment before retrying. The retained pack does not provide the setup procedure.
The saved starting-image reference cannot be loaded.
The graph references two_characters.png, while the interface requires an uploaded Source image and has no uploaded default.
Check: A saved filename does not establish that the corresponding file is available in your environment.
Upload your intended image through Source image so node 16 can load it.
Follow the image, request and returned video
Node 16 (LoadImage) loads the source image. Its IMAGE output at slot 0 connects to image on node 14 (Vidu3ImageToVideoNode). The node schema identifies that image as the generated video's start frame. LoadImage also exposes MASK at slot 1, but it is unconnected here. This recipe has no mask-based region control or compositing stage that establishes original-pixel preservation.
Node 14 receives the image and the generation settings. The exposed controls map as follows:
| Recipe port | Native input | Retained setting or range |
|---|---|---|
source_image | Node 16 image | Upload required; graph references two_characters.png |
prompt | Node 14 prompt | Chase example; empty text allowed; maximum 2,000 characters |
model | Node 14 model | viduq3-pro; also offers viduq3-turbo |
duration | Node 14 model.duration | 6 seconds; integer 1–16 |
resolution | Node 14 model.resolution | 720p; model-dependent options below |
generate_audio | Node 14 model.audio | true |
seed | Node 14 seed | Interface default -1; graph stores 861064363; see seed handling |
The port metadata labels prompt as video and generate_audio as audio, but their bound native inputs are a text string and a boolean. They do not accept an input video or audio recording. This is image-to-video generation, within the generate_video task family. The retained graph does not reveal the provider's internal generation stages.
Node 14's VIDEO output at slot 0 connects to video on node 15 (SweetTeaPreviewVideo). The terminal exposes one video as Generated video, port . Generated sound accompanies that video; there is no separate audio output port. Node 15 retains , and . They are not exposed recipe controls, and their exact encoding behavior is unspecified. The terminal's original-class metadata names ; the actual retained class is , so the graph alone does not establish a saved-file path.
Set length and decide whether the video needs sound
Duration requests the output video's length in whole seconds, from 1 to 16, with a retained value of 6. Prompt text and duration are independent inputs: changing the duration does not rewrite the saved three-shot sequence. Match any requested shot times to the selected length; precise timing adherence is not established.
Generate audio defaults to true and requests sound with the video, including dialogue and sound effects. The prompt describes the desired sounds. Exact dialogue delivery and synchronization are not guaranteed by the retained evidence. Disable the audio flag when the intended output is silent.
- Use six seconds
- Requests a six-second video. Any prompt timestamps must be edited separately to fit.
- Generate silent video
- Disables generated sound in the video. Remove sound directions from the prompt if they no longer describe your intended result.
- Enable generated sound
- Requests sound with the video. Describe desired dialogue and effects in the prompt; this does not provide an exact speech or synchronization guarantee.
Timed prompt instructions extend beyond the selected length.
Duration and prompt text are independent inputs.
Check: Changing Duration leaves the prompt's timestamps unchanged.
Rewrite shot timings to fit, or select a duration within 1–16 seconds that accommodates the intended sequence. Treat those timings as requested guidance.
Sound is requested in the prompt while Generate audio is disabled.
The prompt provides sound directions, while model.audio controls whether generated sound is requested.
Check: Audio is controlled by a boolean input, not by the presence of sound words in the prompt.
Enable Generate audio if sound is required, or remove the sound directions for a silent deliverable.
Choose a compatible model and resolution
Pro supports 720p, 1080p and 2K. Turbo supports 720p and 1080p. The resolution selector lists all three choices, so its option list alone does not establish compatibility with the selected model.
The retained starting pair is viduq3-pro with 720p. The sources provide no measured comparison of Pro and Turbo speed, cost or quality. Choose 2K only with Pro; use a shared resolution before switching to Turbo.
- Use 720p
- Requests a resolution supported by both Pro and Turbo and restores the retained starting resolution.
- Use 1080p
- Requests a resolution supported by both models. No runtime or cost estimate is established.
Which model and resolution configuration do you need?
Turbo is selected while Resolution remains 2K.
Turbo supports only 720p and 1080p.
Check: The selector includes 2K, but the retained interface notes restrict it to Pro.
Select 720p or 1080p, or select Pro if 2K is required, before submitting generation.
Switch to Turbo with a supported resolution
This adaptation uses 1080p, which both models support. Selecting the resolution first avoids leaving a Pro-at-2K configuration paired with Turbo. Submit generation after both settings are selected. This sequence makes no claim that Turbo is faster or cheaper.
- Set Resolution to 1080p before changing the model to Turbo.
Look for: The selected resolution is supported by both Pro and Turbo.
Outputs: output:15
- Select viduq3-turbo, retaining the supported resolution selected in the preceding step.
Look for: The selected pair is Turbo and 1080p, a supported configuration.
Outputs: output:15
Select Pro before requesting 2K
2K is supported with Pro. This sequence selects the model first so that changing resolution does not create a Turbo-at-2K configuration. The pack supplies no exact output dimensions, timing or price for this setting.
- Select viduq3-pro before requesting 2K.
Look for: The selected model supports all three listed resolution choices.
Outputs: output:15
- Set Resolution to 2K while retaining viduq3-pro.
Look for: The selected pair is Pro and 2K, a supported configuration.
Outputs: output:15
Handle the seed default explicitly
The interface documents -1 as randomizing each run and uses it as the displayed default. Both its retained integer constraints and the native node schema specify 0–2147483647, however, while the saved graph stores 861064363. The trailing randomize widget is marked non-runtime; it is not a separate executable input. The retained material does not show where the interface's -1 sentinel is converted before execution. This is a configuration uncertainty, not evidence that randomization fails.
An explicit fixed integer within the native range avoids relying on that unexplained conversion. It supplies a chosen seed but does not guarantee identical provider results. The interface default and saved graph value are distinct records, so seed behavior should not be assumed identical across entry points.
- Use the retained graph's fixed seed
- Explicitly supplies 861064363, which is within the native range. It does not guarantee identical results on later requests.
An entry point rejects the documented randomization value -1 as outside the integer range.
The documented interface sentinel falls outside the retained native nonnegative range; its conversion is not shown.
Check: A fixed integer from 0 to 2147483647 satisfies the recorded range. No observed validation failure is supplied in the pack.
If this occurs, supply a fixed integer in the native range. Use -1 randomization only through an entry point that handles the documented sentinel.
Requests using the same fixed seed produce different videos.
The retained interface explicitly does not guarantee identical provider results.
Check: A fixed seed is a request input rather than an established determinism guarantee.
Record the seed and other request settings when comparing results, and retain the desired video rather than relying on identical regeneration.
Where this approach stops
- Turbo at 2K
- Separate audio output or mask-based region controls in this graph
- Guaranteed identical results from a fixed seed
- Established guarantees of exact shot timing, dialogue synchronization or original-pixel preservation
- Measured runtime, model performance comparisons or hardware-capacity recommendations